The Adversarial Robustness Toolbox (ART) version 1.20.1 contains a flaw that allows an attacker to execute arbitrary Python code during loss function or optimizer evaluation. The vulnerability arises because the evaluation routine uses eval() on user‑supplied strings without sanitization. An attacker who can supply a crafted string will trigger eval() to run the code, giving them full control over the host system where ART runs and resulting in total compromise of confidentiality, integrity, and availability.

The Adversarial Robustness Toolbox (ART) product, specifically the Kubeflow component, is subject to this flaw. Versions up to and including 1.20.1 are affected; later releases may contain a fix. Any environment that integrates ART 1.20.1 or an earlier release and uses the unsafe eval path is at risk.

The vulnerability presents a high‑severity remote code execution risk. Although EPSS data is unavailable, the lack of input validation combined with eval() makes exploitation likely for an entity with access to the Kubeflow interface. The vulnerability is not yet catalogued in CISA’s KEV list, yet deploying the affected ART version exposes the system to full system compromise if a malicious string is supplied during evaluation. The risk remains high until a patch that removes the unsafe eval or a robust input validation strategy is applied.